RAID5 scrub performance

RAID5 scrub performance

[Jorge Bastos]

I believe this is a known issue but wonder if there's something I can
do do optimize raid5 scrub speed, or if anything is in the works to
improve it.

kernel 5.3.8
btrfs-progs 5.3.1


Single disk filesystem is performing as expected:

UUID:             9c0ed213-d9c5-4e93-b9db-218b43533c15
Scrub started:    Tue Nov 26 21:58:20 2019
Status:           finished
Duration:         2:24:32
Total to scrub:   1.04TiB
Rate:             125.17MiB/s
Error summary:    no errors found



4 disk raid5 (raid1 metadata) on the same server using the same model
disks as above:

UUID:             b75ee8b5-ae1c-4395-aa39-bebf10993057
Scrub started:    Wed Nov 27 07:32:46 2019
Status:           running
Duration:         7:34:50
Time left:        1:52:37
ETA:              Wed Nov 27 17:00:18 2019
Total to scrub:   1.20TiB
Bytes scrubbed:   982.05GiB
Rate:             36.85MiB/s
Error summary:    no errors found



6 SSD raid5 (raid1 metadata) also on the same server, still slow for
SSDs but at least scrub performance is acceptable:

UUID:             e072aa60-33e2-4756-8496-c58cd8ba6053
Scrub started:    Wed Nov 27 15:08:31 2019
Status:           running
Duration:         0:01:40
Time left:        1:40:11
ETA:              Wed Nov 27 16:50:24 2019
Total to scrub:   3.24TiB
Bytes scrubbed:   54.37GiB
Rate:             556.73MiB/s
Error summary:    no errors found

I still have some reservations about btrfs raid5/6, so use mostly for
smaller filesystems for now, but this slow scrub performance will
result in multi-day scrubs for a large filesystem, which isn't very
practical.

Thanks,
Jorge

Re: RAID5 scrub performance

[Qu Wenruo]

[-- Attachment #1.1: Type: text/plain, Size: 2077 bytes --]



On 2019/11/27 下午11:11, Jorge Bastos wrote:
> I believe this is a known issue but wonder if there's something I can
> do do optimize raid5 scrub speed, or if anything is in the works to
> improve it.
> 
> kernel 5.3.8
> btrfs-progs 5.3.1
> 
> 
> Single disk filesystem is performing as expected:
> 
> UUID:             9c0ed213-d9c5-4e93-b9db-218b43533c15
> Scrub started:    Tue Nov 26 21:58:20 2019
> Status:           finished
> Duration:         2:24:32
> Total to scrub:   1.04TiB
> Rate:             125.17MiB/s
> Error summary:    no errors found
> 
> 
> 
> 4 disk raid5 (raid1 metadata) on the same server using the same model
> disks as above:
> 
> UUID:             b75ee8b5-ae1c-4395-aa39-bebf10993057
> Scrub started:    Wed Nov 27 07:32:46 2019
> Status:           running
> Duration:         7:34:50
> Time left:        1:52:37
> ETA:              Wed Nov 27 17:00:18 2019
> Total to scrub:   1.20TiB
> Bytes scrubbed:   982.05GiB
> Rate:             36.85MiB/s
> Error summary:    no errors found
> 
> 
> 
> 6 SSD raid5 (raid1 metadata) also on the same server, still slow for
> SSDs but at least scrub performance is acceptable:
> 
> UUID:             e072aa60-33e2-4756-8496-c58cd8ba6053
> Scrub started:    Wed Nov 27 15:08:31 2019
> Status:           running
> Duration:         0:01:40
> Time left:        1:40:11
> ETA:              Wed Nov 27 16:50:24 2019
> Total to scrub:   3.24TiB
> Bytes scrubbed:   54.37GiB
> Rate:             556.73MiB/s
> Error summary:    no errors found
> 
> I still have some reservations about btrfs raid5/6, so use mostly for
> smaller filesystems for now, but this slow scrub performance will
> result in multi-day scrubs for a large filesystem, which isn't very
> practical.

Btrfs uses a not-so-optimal way for multi-disks scrub:
Queuing scrub for each disk at the same time.

So it's common to cause a lot of race and even conflicting seek requests.

Have you tried to only scrub one disk for such case?

Thanks,
Qu

> 
> Thanks,
> Jorge
> 


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 520 bytes --]

Re: RAID5 scrub performance

[Jorge Bastos]

HI,

Thanks for the reply, but I'm not sure I understand, if I start the
scrub for a single device on the raid5 pool it still scrubs the whole
filesystem, and speeds are the same.

Jorge




On Thu, Nov 28, 2019 at 12:01 AM Qu Wenruo <quwenruo.btrfs@gmx.com> wrote:
>
>
>
> On 2019/11/27 下午11:11, Jorge Bastos wrote:
> > I believe this is a known issue but wonder if there's something I can
> > do do optimize raid5 scrub speed, or if anything is in the works to
> > improve it.
> >
> > kernel 5.3.8
> > btrfs-progs 5.3.1
> >
> >
> > Single disk filesystem is performing as expected:
> >
> > UUID:             9c0ed213-d9c5-4e93-b9db-218b43533c15
> > Scrub started:    Tue Nov 26 21:58:20 2019
> > Status:           finished
> > Duration:         2:24:32
> > Total to scrub:   1.04TiB
> > Rate:             125.17MiB/s
> > Error summary:    no errors found
> >
> >
> >
> > 4 disk raid5 (raid1 metadata) on the same server using the same model
> > disks as above:
> >
> > UUID:             b75ee8b5-ae1c-4395-aa39-bebf10993057
> > Scrub started:    Wed Nov 27 07:32:46 2019
> > Status:           running
> > Duration:         7:34:50
> > Time left:        1:52:37
> > ETA:              Wed Nov 27 17:00:18 2019
> > Total to scrub:   1.20TiB
> > Bytes scrubbed:   982.05GiB
> > Rate:             36.85MiB/s
> > Error summary:    no errors found
> >
> >
> >
> > 6 SSD raid5 (raid1 metadata) also on the same server, still slow for
> > SSDs but at least scrub performance is acceptable:
> >
> > UUID:             e072aa60-33e2-4756-8496-c58cd8ba6053
> > Scrub started:    Wed Nov 27 15:08:31 2019
> > Status:           running
> > Duration:         0:01:40
> > Time left:        1:40:11
> > ETA:              Wed Nov 27 16:50:24 2019
> > Total to scrub:   3.24TiB
> > Bytes scrubbed:   54.37GiB
> > Rate:             556.73MiB/s
> > Error summary:    no errors found
> >
> > I still have some reservations about btrfs raid5/6, so use mostly for
> > smaller filesystems for now, but this slow scrub performance will
> > result in multi-day scrubs for a large filesystem, which isn't very
> > practical.
>
> Btrfs uses a not-so-optimal way for multi-disks scrub:
> Queuing scrub for each disk at the same time.
>
> So it's common to cause a lot of race and even conflicting seek requests.
>
> Have you tried to only scrub one disk for such case?
>
> Thanks,
> Qu
>
> >
> > Thanks,
> > Jorge
> >
>

Re: RAID5 scrub performance

[Jorge Bastos]

One last data point on this from a larger pool, 11 disk raid5 pool
(raid1 metadata), this is probably not a very recommended
configuration but it's a 3rd backup and mostly for testing raid5 on a
larger pool:

UUID:             1236acc8-dbd5-41bd-bf3d-872a8fbbce49
Scrub started:    Sun Dec 15 11:23:45 2019
Status:           running
Duration:         0:51:34
Time left:        90:06:52
ETA:              Thu Dec 19 06:22:13 2019
Total to scrub:   18.10TiB
Bytes scrubbed:   175.12GiB
Rate:             57.96MiB/s
Error summary:    no errors found



Same pool during a btrfs send to /dev/null:

btrfs send /mnt/disks/Pics/T12_disk1_Pics_2019-12-15-0829/ | pv -bart
> /dev/null
At subvol /mnt/disks/Pics/T12_disk1_Pics_2019-12-15-0829/
 447GiB 0:13:19 [ 635MiB/s] [ 573MiB/s]
                                   ^                 ^
                             cur. spd   /    avg. spd



So looks to me like scrub performance is around 1/10 of what it should
be, and not optimal to wait multiple days to scrub a larger pool.

Jorge